The most destructive line wasn't a delete
We built a checker that reads the database schema, finds every cascading relation, and fails when a delete guard doesn’t account for one. It found four real bugs in a day. Then I made a completeness claim: there are two schemas, both now loaded; one source directory, now scanned whole. No third schema exists. The boundary is structural, not a judgement call.
It was a good argument. Every fact in it was true.
The falsifying question was one line: what if something deletes without using
the ORM? One grep found DROP DATABASE IF EXISTS on a raw Postgres client in
the provisioning service — the single most destructive statement in the
product, invisible to every version of the scan, because the scan looked for
.delete({ and this wasn’t one.
The site turned out to be safe. It’s a rollback, module-private, one caller, gated on a flag set only after this run created that database itself. Safe and invisible. Invisible was the failure.
What the mistake actually was
Not carelessness — the reasoning inside the frame was sound. The error was that the frame had an unstated edge: destruction goes through Prisma. Never written down, never questioned, and load-bearing for the whole claim.
A raw client dropping a database isn’t a smaller version of a Prisma delete. It’s outside the category the tool was built around. You can’t find it by looking harder at the things you’re already looking at.
The same shape had already bitten us that morning. A guard on deleting an academic term counted classes — sound, careful, and blind to four cascading relations, because it reasoned correctly within “what’s attached to a term” without checking what that phrase excluded.
The rule I’d keep
A completeness claim must name the assumption that would make it false.
Not “I checked everywhere” — here is the thing that, if true, means I didn’t. Mine was “Prisma is how things get deleted.” Stated, it takes one grep to test. Unstated, it survives every audit, because every audit is conducted inside it.
The confidence is the tell. An argument that feels airtight is usually airtight within a frame, and the frame is the part nobody’s looking at.